Skip to content

Draft Model Update

May 24, 2013

I have made significant changes to the draft model; it’s very refined at this point and much improved from the previous iteration. To avoid confusion, I’m going to make this the go-to post and take down all posts pertaining to the earlier versions. But I’m going to use a few quotes from them so they’re not lost.

From my post called “NBA Draft Projection Model and More”:

I’ve been working on developing a model that attempts to project NBA performance of college players. Basically I ran multiple linear regressions on pretty much all the data I could collect (like pace-adjusted box score numbers, team sos, measurements, etc.) for college players from 2002-2009 against their career NBA RAPM. I removed insignificant factors, and eventually came up with a fairly reasonable predictor.

From my post called “Draft Rankings!!”:

I’m obsessive by nature, so this certainly won’t be the final iteration, but I’m pretty happy with it at this point…

…The results are far from perfect as they probably always will be – remember what we’re doing here, we’re taking stats from college kids playing against varying levels of competition in a very limited sample size and trying to project their careers. But I feel pretty confident that we can make educated guesses with this data – and do a much better job than what we’ve actually seen in the past.

I made a number of changes, but the following is a brief summary of the most significant ones:

  • I added in all players’ career numbers instead of using only their numbers from their final NCAA season. This turned out to be a very significant improvement, and I probably should have done it in the first place.
  • I went back to using ONE regression rather than three different ones (for points, wings, and bigs – an idea I had taken from Hollinger). I was able to do this because of having the career numbers. This is important because all players can be more reasonably compared to each other regardless of position, and now everyone is on a sliding scale – for example, the difference between someone who plays a bit more SF than PF and someone who plays a bit more PF than SF is very small now, which is the way it should be.
  • I refined and normalized the y values (the dependent variable for each player). Instead of using long-term RAPM, I used a RAPM-SPM blend based on the same relative period of time for each player.

Here are some observations about what player projections mean:

  • Like before, players +2 or better are very likely to be all-star caliber NBA players. The +2 club is more exclusive than before, and so if there’s a +2 on the board when your team is picking, take him.
  • +1 or better means the player is very likely to be a solid NBA player. In some cases – usually if the player has behavioral or work ethic issues – +1s won’t pan out, but more often than not, they will.
  • Players in the positive or slightly negative range are more likely than not to be solid contributors.
  • Players more than slightly negative will be hit and miss. You probably won’t find may great players here, and the more negative, the less likely it is the player will be any good.

Finally, here are all the out-of-sample (2010 to 2012) results for drafted players:

And as always, the current draft rankings can be found at the top bar.

-James

About these ads
14 Comments leave one →
  1. Nathan permalink
    May 24, 2013 7:41 pm

    This is incredibly accurate…Will you continue posting things like this after you get hired by an NBA front office?

  2. Nathan permalink
    May 24, 2013 10:42 pm

    Jokes aside, I like many others was surprised to see MCW rate so high. Is there any case of such a poor college scorer having success in the NBA? I guess he does pretty much everything but score at an elite level, though. Exciting prospect if you think you can get him to buy into being a pure facilitator.

  3. ljv permalink
    May 25, 2013 5:20 am

    “I refined and normalized the y values (the independent variable for each player). Instead of using long-term RAPM, I used a RAPM-SPM blend based on the same relative period of time for each player.”

    Do you mean “the *dependent variable* for each player*?

    I agree that it is better not to run separate models for each position or position group. DO you still include position as a predictor in the model or as an interaction with any of the other predictors?

    I’m also curious what combine measures added value. I found that only standing reach and no-step vert survived, but I wouldn’t be surprised if there are ways to mlk information from some of the others.

    • Nathan permalink
      May 25, 2013 6:46 am

      According to Kevin Hetrick’s work at Hardwood Paroxysm the agility drill has some predictive value.

    • May 25, 2013 7:55 am

      Oops, good catch ljv. Yes i include position, and yes some interaction with other predictors (rebounding is the big one). The only combine factors i have found to be significant are wingspan and height.

      You said you used standing reach and no-step vert post-hoc, correct? I might have to check that out, but I’d be surprised if anything else had any real value.

      • ljv permalink
        May 25, 2013 10:50 am

        Interesting, I found that including standing reach made both height and wingspan irrelevant… though I can see why wingspan may still carry important information for guards in particular.

        No-step vert has consistently been the best physical measure in the different ways I approached this. When I was only using the DX dataset and thus had combine info for most guys, I included physicals as predictors as you are. No-step vert was the best predictor by quite a bit in that scenario (I think standing reach and weight were the only others to hang around). Now that I am working with the BR dataset, I can’t include physicals because there aren’t any for guys from the 80s and 90s, so I did it post-hoc on the results.. so basically re-predicting the same y but with the post-2000 dataset with the step-1 prediction and physicals as predictors. The only way I got anything to improve the story was making no-step vert an interaction effect with step-1 prediction.. such that vertical matters for guys who put up good numbers, but not for guys who don’t.

  4. Nathan permalink
    May 26, 2013 4:55 pm

    Noel’s injury obviously has an impact on his training, if nothing else. For the 9 months he’s out, maybe he only makes 1/3 the improvement he would have made if healthy.

    If you have time, could you run the rater on Noel with his age set 6 months older than he actually is? I’m curious to know if he still rates as the best prospect, and if so by how much.

    • May 27, 2013 11:57 am

      Nathan,

      Very interesting idea. At six months older, Noel still rates as the best player in the draft by a substantial margin.

  5. Safwaan permalink
    June 22, 2013 10:34 pm

    Is Brandon Knights absence in your 2011 rankings due to a low score or did you not input his data

    • June 25, 2013 7:32 am

      Safwaan,

      Good catch. I’m not sure how I excluded Brandon Knight, but he’s in there now.

  6. Brad permalink
    June 9, 2014 1:03 pm

    This is very impressive, I had a question about the following statement:

    “I added in all players’ career numbers instead of using only their numbers from their final NBA season.”

    Do you mean that you included a players career NCAA numbers in calculating the predictor (x) variables? Or are you saying that you included a players career NBA numbers for calculating their RAPM-SPM?

    • June 9, 2014 4:52 pm

      I think that should read “final NCAA season”…e.g., I used all ncaa seasons instead of the final season only, which is what I had used in previous versions of the model…I’ll fix that.

      • Brad permalink
        June 10, 2014 9:06 am

        Thanks James,

        I find your model very intriguing. I’m curious if you weighted all of a players NCAA seasons equally or gave additional weight to the most recent seasons an less weight to theri earlier seasons?

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

Follow

Get every new post delivered to your Inbox.

%d bloggers like this: